一种基于YOLOv8模型的轻量级密集行人检测算法  

A lightweight dense pedestrian detection algorithm based on YOLOv8 model

作　　者：赵艳芹 王璐瑶 ZHAO Yanqin;WANG Luyao(School of Computer and Information Engineering,Heilongjiang University of Science and Technology,Harbin 150022,China)

机构地区：[1]黑龙江科技大学计算机与信息工程学院,黑龙江哈尔滨150022

出　　处：《高师理科学刊》2025年第2期32-39,共8页Journal of Science of Teachers'College and University

基　　金：黑龙江省省属本科高校基本科研业务费项目(2022-KYYWF-0565)。

摘　　要：针对高密度行人检测中的遮挡、小目标漏检及算法参数量大的问题,基于YOLOv8模型提出了一种轻量级行人检测算法,称之为YOLOv8-DSR算法.首先,在骨干网络引入改进的可变形卷积块,用以增强网络对不规则行人的特征提取能力;其次,在原YOLOv8模型基础上新增160×160尺度的检测头,降低了小尺度行人漏检率;最后,引入轻量化卷积GhostConv模块及全新的C2f_RepGhost模块,降低了特征融合过程中的运算量,减少了模型参数量,节省了硬件推理时间.为验证改进的有效性,在CrowdHuman和WiderPerson数据集上进行实验,结果表明,相较于YOLOv8n算法,YOLOv8-DSR算法参数量减少了18.75%,在检测精度方面,YOLOv8-DSR算法在CrowdHuman数据集上的mAP@0.5和mAP@0.5~0.95分别提升了3.6%,2.6%,在WiderPerson数据集上的mAP@0.5和mAP@0.5~0.95分别提升了2.1%,2.3%.改进后算法的参数量更小,检测精度更高,对硬件的要求更低,更便于模型部署.In response to the challenges of occlusion,missed detection of small targets,and large algorithmic parameter counts in high-density pedestrian detection,a lightweight pedestrian detection algorithm based on the YOLOv8 model is proposed,named the YOLOv8-DSR algorithm.Firstly,an improved deformable convolution block is introduced into the backbone network to enhance the network′s ability to extract features from irregular pedestrians.Secondly,a detection head with a 160×160 scale is added to the original YOLOv8 model to reduce the miss-detection rate of small-scale pedestrians.Finally,the lightweight convolution GhostConv module and the novel C2f_RepGhost module are incorporated,which reduce the computational load during feature fusion,decrease the model′s parameter count,and save hardware inference time.To verify the effectiveness of these improvements,experiments are conducted on the CrowdHuman and WiderPerson datasets.The results show that compared with the YOLOv8n algorithm,the proposed YOLOv8-DSR algorithm reduces the parameter count by 18.75%.In terms of detection accuracy,the YOLOv8-DSR algorithm achieves a 3.6%increase in mAP@0.5 and a 2.6%increase in mAP@0.5~0.95 on the CrowdHuman dataset,and a 2.1%increase in mAP@0.5 and a 2.3%increase in mAP@0.5~0.95 on the WiderPerson dataset.The improved algorithm has fewer parameters,higher detectionaccuracy,and lower hardware requirements,making it more convenient for model deployment.

关 键 词：行人检测 YOLOv8 可变形卷积 轻量化 小目标 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]